Research On Text Sentiment Analysis Based On ALBERT SABL Model

With the rapid iteration of the Internet,the era of big data has come.People are increasingly accustomed to using the Internet to get the information what they want.People also output their personal views on the Internet,such as the opinions of events and comments about the certain products.Through these subjective data can dig out the emotional tendency of the public,sentiment analysis is diffusely used in e-commerce comments and online public opinion,has important research value.This summarizes the crucial technique of text sentiment analysis and finds that the method is commonly used the framework combining dictionary and deep learning technology,which can perform well in the baseline task.As time goes by,the lack of contextual semantic ability leads to gradient disappearance.With the increasing layers of deep neural network,the parameters increase exponentially,that leads to the model structure more complex and affects the speed of training.In view of the appeal problem,this paper proposes a text sentiment analysis method based on ALBERT-SABL model.The specific research content is summarized as follows:Firstly,for the problem of gradient disappearance,analyzes the traditional attention mechanism algorithm,improves on this basis,proposes a new Hybrid self-attention mechanism algorithm,which uses the combination of soft and hard attention to build a Hybrid matrix of bidirectional information enhancement.To increase the weight of important information.Secondly,in terms of training effect,ALBERT pre-training model based on context association is proposed with fewer parameters,BERT's Masked LM task is modified on the original basis to make it more compatible with text emotion classification task and more effective in learning sentence representation.Finally,the above two algorithms are integrated Bi LSTM model into ALBERT-SABL model.This method uses the ALBERT model to train the word vector as the word vector embedding layer,uses the bidirectional neurons in the hidden layer to construct a semantic hierarchical structure between contexts,trains the input word vector,extracts the data features of the text,and then inputs it into the Hybrid self-attention mechanism computes internal word dependencies.Experiments on Restaurant,Laptop and Twitter data sets,show that ALBERT-SABL model has better performance than other traditional models,which proves that this model has certain research significance and application value in the field of emotion analysis.